⌁ Awesome AI4CAD Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Self-Improving CAD Generation Agents with Finite Element Analysis as Feedback

arXiv 2026 65.9 method

TLDR

Introduces self-improving CAD agents that use finite element analysis feedback to generate multi-part STEP files from engineering briefs, improving geometric reconstruction.

Reasoning

Strengths include a novel industry-native task formulation and the use of FEA as a practical validation signal, moving beyond geometric similarity. Weaknesses are limited evaluation on only two datasets and modest absolute performance gains, with no real-world deployment or user study.

Read-first score

Read-first score 65.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 77.

Methodology quality 25%
100

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,dataset,evaluation,metric,validation

Recency 8%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Reproducibility 25%
50

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=artifact,code,dataset

Topical relevance 42%
48.1

Uses existing LLM keyword relevance scores normalized to 0-100. AI for CAD,computer-aided design,neural CAD,generative CAD,parametric CAD,B-Rep,boundary representation,constructive solid geometry,CSG,sketch extrusion,CAD generation,CAD reconstruction,text-to-CAD,image-to-CAD,point cloud to CAD,CAD program

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 13.

Keyword Scores

AI for CAD
9
computer-aided design
9
CAD generation
9
generative CAD
8
text-to-CAD
8
CAD reconstruction
7
neural CAD
6
CAD program
6
parametric CAD
4
image-to-CAD
3
B-Rep
2
boundary representation
2
constructive solid geometry
1
CSG
1
sketch extrusion
1
point cloud to CAD
1

Deep Analysis

Innovations

  • Introduction of an industry-native task formulation requiring generation of fully assembled multi-part STEP files from free-form engineering briefs, validated via finite element analysis (FEA).
  • Integration of FEA as a feedback signal to enable self-improving CAD generation agents, moving beyond geometric proximity metrics.
  • Two novel supervision signals: a text-only blueprint schema and a 21-view image renderer for visual inspection, aligning the generation loop with engineering iteration practices.

Methodology

The paper proposes a task where agents (Codex GPT-5.5, Claude Opus-4.7) generate multi-part STEP files from engineering briefs, then validate outputs using FEA. Two additional feedback tools—a text blueprint schema and a 21-view image renderer—are introduced to guide iterative refinement. Evaluation is performed on S2O and Fusion360 datasets using Box-IoU and requirement satisfaction metrics.

Key Results

FEA validation shows no strict-passing artifact in first-attempt sweeps; the best configuration meets only ~20% of typed requirements on average. With the proposed feedback tools, GPT-5.5/xhigh Box-IoU improves from 0.444 to 0.592 on S2O and from 0.397 to 0.505 on Fusion360.

Limitations

  • Agents fail to produce any strict-passing artifact in first-attempt FEA validation.
  • Best configuration meets only about 20% of typed requirements on average, indicating low overall compliance.
  • The feedback signals improve geometric reconstruction (Box-IoU) but the abstract does not report achieving strict FEA passing even after iteration.
  • Evaluation is limited to two datasets (S2O, Fusion360) and two proprietary agent models.

Resources

Paper

Tags

GRCL